Your data
Load a dataset and Mizura can work with it. This page lists what you can load, how to import it, how to inspect it, and where the limits are.
Supported formats
Mizura reads tabular files and works out the layout from the file itself. Delimited text gets its delimiter detected. A file is accepted only if it yields more than one column and fewer than half the columns are entirely empty.
| Kind | Extensions | Notes |
|---|---|---|
| Delimited text | .csv .tsv .tab .txt .dat | Delimiter detected from the file. Whitespace is the last resort. |
| Spreadsheets | .xlsx .xlsm .xls .ods | The dataset is the first sheet. .xlsx and .xlsm work out of the box. An older .xls or .ods that fails to read needs converting to .xlsx or CSV first. |
| Columnar | .parquet .pq .feather | Read directly. |
| JSON | .json | JSONL is not a dataset format. |
| Stata | .dta | Read directly. |
| SPSS | .sav .zsav | Read directly, including labelled variables and missing-value metadata. |
| SAS | .sas7bdat .xpt | Read directly. A separate .sas7bcat is needed to interpret custom SAS value labels. |
JSONL is not supported as a dataset. Other extensions are rejected with the supported list. R, Stata, SPSS, and SAS also appear on the output side: Mizura can write reference code in those languages for a finished Python analysis, but does not execute that code. See How code runs.
A corrupt file fails with an error that says what was tried. Nothing is left on disk.
If the upload is rejected, check the extension and size limits below. If it is unreadable, re-export or convert it to a supported tabular format, then upload it again. Confirm the dataset preview and column names before continuing.
Size limits
| Limit | Default | What happens past it |
|---|---|---|
| Direct upload, streamed formats (delimited text, Parquet, Stata, SPSS, and SAS) | 10 GB | Upload rejected. |
| Direct upload, in-memory formats (spreadsheets, Feather, and JSON) | 200 MB | Upload rejected. |
| Imported files (GitHub, Drive, Zenodo) | 200 MB | Import rejected before anything is written. |
| Large-dataset threshold | 100 MB or more, streamed format | Handled without loading the whole file into memory. See below. |
| Attachments per message | 5 | Further attachments are refused for that message. |
Large datasets
A streamed dataset of 100 MB or more is never loaded whole. Mizura reads it in place, and the preview and profile come from the file itself. The profile uses a repeatable 50,000-row sample and says so. Treat those numbers as indicative until the analysis runs.
Replication packages leave large datasets out. The package records the dataset's identity and explains how to rebuild it. See Reports & exports.
Inspect before you analyze
Loading gives you three views. Use them. Thirty seconds here saves an hour of debugging a wrong-variable analysis later.
- Preview. The first rows, in the dataset panel. Check the column names. Check that the first rows are not misleading, for example a blank first value.
- Summary statistics. Count, mean, standard deviation, min, quartiles, median, max, skew. Numeric columns only.
- Full inspector. Open the dataset chip for the complete table (sortable, searchable, paginated), the summary statistics, and the column list with types. The header shows row and column counts plus the missing-value count.
The agent's profile goes further. It covers the first 150 columns: numeric ranges and central values, or the top five values for anything else. It also flags facts worth knowing. Sentinel codes like -9999 sitting outside the rest of the range. Text placeholders like NA that pandas counts as present. Numbers stored as text. Labels that differ only by case or spacing. Constant columns. Identifier columns unique per row. Duplicate rows. These are observations, not verdicts. They tell you what to check.
Import from GitHub, Google Drive, and Zenodo
You can pull a dataset instead of uploading one. Importing ends in the same place as uploading: the file is stored, previewed, and pinned to its source.
| Source | What you do | Sign-in |
|---|---|---|
| Zenodo | Paste a DOI (10.5281/zenodo.123) or a record URL, browse the files, import one. | None. Public records only. |
| GitHub | Paste a project URL. Browse dataset files at the latest commit (up to 500 listed). Import one at its pinned commit. | Optional. Public projects need none. Connect GitHub in Settings for private projects. |
| Google Drive | Search your Drive for dataset files. Import one. Sheets come in as CSV. | Required. Connect Google Drive in Settings, read-only. The file must be a supported format or a Sheet. |
Imports check the size cap before, during, and after the download. Connections use secure hosts only, and credentials are dropped if a download redirects elsewhere. Connector connections live on your device. Mizura never displays a connection token anywhere.
For the steps, connection requirements, and what gets stored, see Import from connected services.
Attachments and sources are not datasets
Two other kinds of file enter a session. Neither becomes your dataset.
- Message attachments. Up to five per message. Code, documents, spreadsheets, images. The agent reads them in sections: text and code directly, PDF and DOCX extracted to text, spreadsheets as tables. Images cannot be read as text. A scanned PDF yields nothing, and the agent is told so instead of guessing.
- Files & Context sources. PDFs, text files, Markdown, and URLs you add in the left panel. Codebooks. Questionnaires. Papers your analysis should know about. They show up in the Sources tab next to the literature the agent finds itself. See Evidence & literature.
Changing the dataset mid-session
The agent sometimes prepares a new version of your dataset as part of an analysis. A filter. A recode. When that happens a dataset-change card appears in the dataset panel: what changed, how many rows were removed, added, or edited, and whether anything was flagged. Undo it. Download the previous data. Read the details. Changes are recorded and reversible. The change is applied when the analysis run saves it; Mizura does not ask for a separate approval first. If a run is in progress, undo is available after it finishes.